Input sanitization cleans and validates user inputs before LLM processing to prevent attacks. Purposes: Block prompt injection attempts, filter harmful content, normalize inputs, validate format. Techniques: Keyword filtering: Block known attack patterns ("ignore previous", "system prompt"). Encoding detection: Flag base64, hex, or obfuscated text that may hide payloads. Length limits: Prevent prompt stuffing attacks. Character filtering: Remove or escape special characters, control codes. Format validation: Ensure expected input structure (JSON, specific fields). Content scanning: Check for toxic content, PII, code injection. Limitations: Adversarial inputs constantly evolve, over-filtering harms usability, semantic attacks bypass keyword filters. Layered approach: Input sanitization + system prompt design + output filtering + monitoring. Implementation: Pre-processing pipeline before LLM call, can use regex, classifiers, or another LLM as detector. Best practices: Allowlist over blocklist, defense in depth, log flagged inputs, regular pattern updates. Essential first layer of defense but not sufficient alone.